SINA: accurate high throughput multiple sequence alignment of ...

Supplemental Material

SINA: accurate high throughput multiple sequence alignment of ribosomal RNA genes Elmar Pruesse 1,2 , J¨org Peplies 3 Frank Oliver Gl¨ ockner 1,2

1

Max Planck Institute for Marine Microbiology, Bremen, Germany, Jacobs University Bremen, Bremen, Germany 3 Ribocon GmbH, 28359 Bremen, Germany 2

1 1.1

Algorithm Positional Variability by Parsimony (PVP)

The “positional variability by parsimony” (PVP) function of ARB computes a percolumn conservation profile from a MSA and a phylogenetic tree. For each column, the number of transitions and transversions required to explain the tree given the aligned sequence data is computed. The sum of transitions and transversions divided by the number of observed bases, capped at 0.55 and corrected for not observed mutations using the Jukes Cantor formula (Jukes and Cantor, 1969). From this rate, we compute the scoring weight as 0.5-log(rate). This weight is capped at 20. Columns containing gap characters in more than 80% of the sequences are assigned a weight of 1. The PVP statistic applied may be chosen dynamically for each candidate sequence based on classification meta-data in the reference MSA database. This allows using for example domain specific statistics. The name of each PVP statistic stored in the reference database is compared with the configured classification attribute of each reference sequence. If a name is a prefix of the attribute for a majority of the sequences, the corresponding PVP statistic is used instead of a globally configured PVP statistic.

2

Results

All figures displaying mean accuracy (all except Fig. S2) are shown twice, once using a linear (labeled A) and once using a logarithmic scale on the y-axis (labeled B).

References Jukes, T. H. and Cantor, C. R. (1969). Evolution of protein molecules. New York: Academic Press, pages 21–132.

1

99% 98% 95%

A

mean accuracy (Q score)

90%

number of reference sequences 80%

1 2 5 10

70%

20 40 100

60%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

99%

B

mean accuracy (Q score)

98%

number of reference sequences 1 2

95%

5 10 20

90%

40 100

80%

70% 60%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S1: Effect of increasing the number of reference sequences on alignment accuracy at different levels of identity with the reference alignment.

2

7%

bases w/o column assignment (insertions)

6%

number of reference sequences

5%

1 4%

2 5 10

3% 20 40 100

2%

1%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S2: Effect of increasing the number of reference sequences on the fraction of “insertions” with respect to the alignment template.

3

99% 98%

A

mean accuracy (Q score)

95%

match score 2

90%

1.5 1

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

B

mean accuracy (Q score)

99%

98% match score 2 1.5 1

95%

90%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S3: Using a mismatch score of -1 a match score of 2 is minimally better than a match score of 1 or 1.5. The difference, however, is almost beyond the resolution of this figure.

4

99% 98%

A

mean accuracy (Q score)

95%

gap penalties (open/extend) 5/2 90%

5/3 5/5 2/1 2/2

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

B

mean accuracy (Q score)

99%

gap penalties (open/extend)

98%

5/2 5/3 5/5 2/1 95%

2/2

90%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S4: A gap open penalty of 5 works better than a gap open penalty of 2. Among the tested gap extend penalties using a gap open penalty of 5, a gap extension penalty of 2 is best by a small margin.

5

99% 98%

A

mean accuracy (Q score)

95%

configured full length none 90%

1300 1400 1500 1600

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

B

mean accuracy (Q score)

99%

configured full length

98%

none 1300 1400 1500 1600

95%

90%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S5: Requiring that at least one sequence of 1500 or 1600 bp length be included in the reference set improves averaged results.

6

99% 98%

A

mean accuracy (Q score)

95%

method of fixed column preservation

90%

base shifting DP insertion limiting

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

B

mean accuracy (Q score)

99%

98% method of fixed column preservation base shifting DP insertion limiting

95%

90%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S6: The method used for column preservation makes little difference to alignment accuracy using the SSU benchmark. Considering that the SSU alignment is more than 30 times wider than typical SSU sequences and that special care was taken to have sufficient alignment space between bases in the construction of the alignment, this is not unexpected.

7

99% 98%

95%

A

mean accuracy (Q score)

k−mer length 6 7 8

90%

9 10 11 12

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

99%

B

mean accuracy (Q score)

98%

k−mer length 6 7 8 9

95%

10 11 12

90%

80%

70%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S7: Varying the length of the kmeres used to find reference sequences has little impact, values between 8 and 12 work well for SSU sequences.

8

99% 98%

A

mean accuracy (Q score)

95%

search type normal 90%

fast normal+full fast+full

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

B

mean accuracy (Q score)

99%

98% search type normal fast normal+full fast+full 95%

90%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S8: Using the “fast mode” of the kmer search provided by the ARB PT server ignores kmeres that do not begin with ’A’. The impact to alignment accuracy is visible. The graphs labled “+full” require that at least one sequence of at least 1400 bp be included in the reference.

9

99% 98%

A

mean accuracy (Q score)

95%

weighting scheme weight=1 generic PVP 90% domain PVP generic PVP + weight=2 generic PVP + weight=5 unweighted

80%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

B

mean accuracy (Q score)

99%

weighting scheme 98% weight=1 generic PVP domain PVP generic PVP + weight=2 generic PVP + weight=5 95%

unweighted

90%

40%

50%

60%

70%

80%

90%

100%

mean identity with closest reference sequence

Figure S9: All weighting schemes and combinations thereof provide some improvement to unweighted alignment. Using only the base frequency among the selected reference sequences performs significantly better than all other tested methods.

10